← MCP server and skill grades

skills

caffeinelabs/skills/skills

B
Type
Skill
Scanned
Nov 13, 2026
Version
59b9b29a4fad
Files read
59

This project provides AI coding agent skills and a local CSV ingestion script for building full-stack Motoko and React applications on the Caffeine Internet Computer platform. The main risks are that the scaffolding skill instructs the agent to run unpinned npm packages without confirmation and fetch live remote skill instructions at runtime, and several connector and email recipes contain OAuth, authentication, or header-injection flaws that can be copied into generated apps.

What it can reach

Files
The bundled oql-ingest script reads local CSV files and writes temporary Candid argument files in the OS temporary directory, and the skills guide creating and editing project source files.
Shell commands
The oql-ingest script executes the icp CLI binary, and the skills instruct the agent to run npm, npx, pnpm, mops, caffeine, icp, and curl commands.
Network
The oql-ingest script sends canister calls via icp, and the skills instruct fetching packages and remote skill files from npm, mops, github.com, caffeine.ai, and skills.internetcomputer.org, plus probing external APIs with curl.
Secrets
The skills guide authenticating the caffeine CLI (storing tokens in the OS keychain or ~/.caffeine/tokens.json), using icp controller identities for bulk data uploads, and wiring third-party API keys and OAuth secrets into canisters.
Browser
The caffeine-app skill instructs running CLI commands such as caffeine auth login and caffeine chat open that launch the user's browser to caffeine.ai.

Findings

2 medium1 low
  1. MediumSkill instructs running unpinned npm packages with auto-confirmation

    skills/caffeine-app/SKILL.md:68

    The caffeine-app skill instructs the coding agent to execute npx -y @caffeineai/cli@latest (as well as unpinned npm install -g @caffeineai/cli and npx skills experimental_install -y). Using -y with @latest or unpinned package names downloads and executes whatever version is currently published on npm without confirmation, and experimental_install -y automatically installs third-party skills listed in a cloned project's skills-lock.json.

    Fix: Pin exact verified package versions instead of @latest, remove -y auto-confirmation flags for package and skill installations, and instruct users to review skills-lock.json before installing component skills.

  2. MediumSkill directs agent to fetch and obey unpinned remote instructions at runtime

    skills/caffeine-app/SKILL.md:20

    Even though the writing-motoko skill and component skills are included locally in this package, caffeine-app instructs the agent to fetch and follow remote markdown files over HTTPS (such as https://skills.internetcomputer.org/skills/writing-motoko/SKILL.md and https://github.com/caffeinelabs/skills) before editing backend code. Dynamically loading unpinned remote instructions bypasses static review of the installed skill and exposes the agent to remote prompt injection if the remote host or repository is compromised.

    Fix: Reference the locally installed writing-motoko and component skill files via relative paths rather than fetching live SKILL.md files from external URLs at runtime.

  3. LowMandatory X OAuth recipe omits state parameter verification (OAuth CSRF cattle-link risk -> OAuth CSRF account-linking risk) -> wait let's fix typo!

    skills/connector-x/SKILL.md

    ...